The documents distributed by this server have been provided by the contributing authors as a means to ensure timely dissemination of scholarly and technical work on a noncommercial basis. Copyright and all rights therein are maintained by the authors or by other copyright holders, notwithstanding that they have offered their works here electronically. It is understood that all persons copying this information will adhere to the terms and constraints invoked by each author's copyright. These works may not be reposted without the explicit permission of the copyright holder.
Publications of SPCL
|B. Prisacari, G. Rodriguez, P. Heidelberger, D. Chen, C. Minkenberg, T. Hoefler:|
|Efficient Task Placement and Routing in Dragonfly Networks |
(. Vol , Nr. , In Proceedings of the 23rd ACM International Symposium on High-Performance Parallel and Distributed Computing (HPDC'14), presented in Vancouver, Canada, pages , ACM, ISSN: , ISBN: , Jun. 2014, )
AbstractDragonflies are recent network designs that are one of the most promising topologies for the Exascale effort due to their scalability and cost. While being able to achieve very high throughput under random uniform all-to-all traffic, other widely used workloads in high performance computing, such as the Cartesian nearest neighbor communication pattern, experience a significant performance degradation, and have been termed as ‘pathological’ patterns for Dragonfly networks. Despite the importance of the Cartesian nearest neighbor pattern as a basic building block for most scientific applications, there is as of now, no detailed theoretical model which can predict its performance accurately on a dragonfly topology taking into account (i) the network parameters (number of links and bandwidths per level) (ii) the routing function used in the network (direct/indirect) and (iii) the placement on the network (random, hierarchical, etc.). Such a theoretical model would establish the real performance expectation for this communication pattern, as well as help determine the parameters that would lead to an optimum performance. In this work we present such a theoretical performance analysis framework, which, taking as input (i) a network specification, (ii) the domain decomposition parameters from which the nearest neighbor communication pattern emerges, and (iii) a task placement description, can predict the impact on the effective sustained injection bandwidth. Our analysis is able to provide analytic bounds for the peformance of the neareat neighbor exchange across a multitude of possible task-to-node maping strategies. Finally, using a comprehensive set of simulations results, we validate the correctness of the theoretical approach and in the process address some misconceptions regarding Dragonfly network behavior and evaluation, (such as the choice of throughput maximization over workload completion time minimization as optimization objective) and the question of whether the standard notion of a balanced Dragonfly (topological parameters chosen for near-optimum performance for uniform random all-to-all traffic) can be extended to workloads other than uniform random traffic. To this end we conclude that the optimal performance/cost Dragonfly network design for uniform random all-to-all traffic is different to the optimal performance/cost Dragonfly network design for the Cartesian nearest neighbor pattern.